The Agent Skills Directory

COMMAND_EXECUTION (SAFE): The skill is designed to facilitate the execution of developer-focused CLI tools (gemini, claude, opencode, codex) via bash. While it uses high-privilege modes like pty:true and background:true, these are standard for interactive terminal applications and align with the skill's stated purpose of programmatic coding assistance.
PROMPT_INJECTION (LOW): Indirect Prompt Injection surface. The skill ingests untrusted data from both the user (via command strings) and the local environment (via workdir and file access) and passes it to external LLM-powered agents.
Ingestion points: The command parameter in bash calls and the files processed within the specified workdir (e.g., ~/project).
Boundary markers: No specific delimiters or "ignore instructions" warnings are utilized in the provided examples to separate system instructions from processed data.
Capability inventory: The skill allows for full bash execution, pseudo-terminal (pty) access, and background process management across multiple coding agents.
Sanitization: No input sanitization or validation of the commands or code being processed is performed by the skill itself.
EXTERNAL_DOWNLOADS (SAFE): The skill's metadata requires specific binaries (claude, codex, opencode, gemini, pi). These are references to well-known tools from trusted organizations (Google, Anthropic, Microsoft) and the skill does not attempt to download them from untrusted sources at runtime.

ogt-cli-agent